Pull request status on the panel

When the branch an agent is working on has an open pull request, its number appears in the panel header. The color shows the result of the automated checks: green when they passed, red with a count when one failed, and in progress while they are still running. The same indicator appears in the footer and in the Git panel for the current branch.

Actions from the indicator

A click opens the list of checks, with how long each one took, and three options: open the pull request in the built-in browser, view the failure log, or ask the responsible agent for a fix. In the last option, the agent receives the failure details together with the request.

Immediate failure notice

When a check fails, the notice appears right away and states which agent and what broke, even with another project in the foreground.

Separate browser sessions

What agents open in the browser uses a session separate from yours. You can stay signed in to services such as GitHub in the canvas browser without agents having access to that session.

Turning all of it off

In Settings, a single option removes the indicator, the notices, and the requests to GitHub.

Performance panel

A new panel, opened with ⌘K and terms such as performance, slow, or memory, shows memory and processor usage by name. Agents, browsers, and terminals are listed separately, so it is clear what is consuming the most.

Automatic report when usage is unusual

The application learns how much it usually consumes in your workflow. When usage departs from that pattern, it records a report of the moment: open agents, what each one was doing, browsers, and the minute-by-minute increase. The report stays on your machine.

Status bar indicator

Next to the zoom control there is a discreet number that takes on color when usage is unusual. Clicking it opens the performance panel.

Sensitivity setting

In Settings, under Performance, the only choice is how large a deviation should raise a notice. There is no gigabyte field: the reference is measured by the application itself and follows your usage.

Idle agents can sleep

When enabled in Settings, under Agents, an agent with no activity for a period goes to sleep and releases memory. The panel then shows a Wake button, and the conversation resumes where it stopped. Agents that are working, agents waiting for an answer, and the panel in focus never sleep. Sending a message, by text or by voice, also wakes the agent, which receives it as soon as it is back. The feature ships disabled, and the application points to it when idle agents are consuming memory.

External tools chosen per canvas

When creating or editing a canvas, you can choose which external tools are available to the agents of that project instead of loading all of them. In internal measurements, usage per agent dropped from 482 MB to 190 MB. If the choice is changed with agents open, they come back with the new configuration without losing the conversation.

Current topic for each agent

Below the name, in the panel, the agent itself describes in one line what it is working on: Payment integration tests, Memory leak in the player. It is not an excerpt of the request or of the answer, but the topic in plain language. A glance at the canvas shows where every agent stands.

Topic set by hand

A double click on the line allows writing any text. From then on no agent changes it. To hand control back to the agent, clear the text or use the matching option in the context menu.

Redesigned panel back

The back of the panel no longer lists action counts. It now shows what the agent wrote about the current work: the topic, the context in two or three jargon-free sentences and, highlighted, what it is waiting for from you.

More accurate voice summary

When Nina is asked how the agents are doing, the answer now uses these descriptions, including who is idle waiting for an answer and what they need to know.

Header follows the interface scale

The agent name and the topic line follow the interface scale set in Settings.